Blinded Predictions and Post Hoc Analysis of the Second Solubility Challenge Data: Exploring Training Data and...

Area of Science:

• Chemical Sciences
• Computational Chemistry
• Drug Discovery

Background:

• Accurate prediction of solubility from molecular structure is a significant challenge in chemical sciences.
• The American Chemical Society's "Second Solubility Challenge" (2019) aimed to assess the state-of-the-art in solubility prediction.
• Competitors submitted blinded predictions for 132 drug-like molecules.

Purpose of the Study:

• To develop and report two novel machine learning models for solubility prediction submitted to the 2019 challenge.
• To evaluate the impact of advanced algorithms and larger datasets on prediction accuracy by comparing original models with deep learning models.
• To analyze the performance differences between models based on feature sets and training data.

Main Methods:

• Development of two traditional machine learning models using inexpensive molecular descriptors and a small training set (300 molecules).
• Training and evaluation of deep learning models, including graph convolutional neural networks, on larger datasets (2999 and 5697 molecules).
• Comparison of prediction accuracy (RMSE) between original and deep learning models on the challenge dataset.

Main Results:

• Several algorithms achieved near state-of-the-art performance.
• A graph convolutional neural network model yielded the best performance with an RMSE of 0.86 log units.
• Systematic differences in model performance were observed based on feature sets and training data quality and relevance.

Conclusions:

• Machine learning models can achieve high accuracy in predicting solubility.
• The quality and relevance of training data are critical for accurate solubility prediction.
• Challenges remain in modeling complex chemical spaces with sparse training data, highlighting the need for careful data selection and methodological improvements.